1 /*
2  * Copyright (c) 2003, 2007-14 Matteo Frigo
3  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License as published by
7  * the Free Software Foundation; either version 2 of the License, or
8  * (at your option) any later version.
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write to the Free Software
17  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
18  *
19  */
20 
21 /* This file was automatically generated --- DO NOT EDIT */
22 /* Generated on Thu Dec 10 07:05:04 EST 2020 */
23 
24 #include "dft/codelet-dft.h"
25 
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27 
28 /* Generated by: ../../../genfft/gen_notw.native -fma -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n2sv_32 -with-ostride 1 -include dft/simd/n2s.h -store-multiple 4 */
29 
30 /*
31  * This function contains 372 FP additions, 136 FP multiplications,
32  * (or, 236 additions, 0 multiplications, 136 fused multiply/add),
33  * 138 stack variables, 7 constants, and 144 memory accesses
34  */
35 #include "dft/simd/n2s.h"
36 
n2sv_32(const R * ri,const R * ii,R * ro,R * io,stride is,stride os,INT v,INT ivs,INT ovs)37 static void n2sv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38 {
39      DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
40      DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
41      DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
42      DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
43      DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
44      DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
45      DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
46      {
47 	  INT i;
48 	  for (i = v; i > 0; i = i - (2 * VL), ri = ri + ((2 * VL) * ivs), ii = ii + ((2 * VL) * ivs), ro = ro + ((2 * VL) * ovs), io = io + ((2 * VL) * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) {
49 	       V T7, T4r, T4Z, T18, T1z, T3t, T3T, T2T, Te, T1f, T50, T4s, T2W, T3u, T1G;
50 	       V T3U, Tm, T1n, T1O, T2Z, T3y, T3X, T4w, T53, Tt, T1u, T1V, T2Y, T3B, T3W;
51 	       V T4z, T52, T2t, T3L, T3O, T2K, TR, TY, T5F, T5G, T5H, T5I, T4R, T5k, T2E;
52 	       V T3M, T4W, T5j, T2N, T3P, T22, T3E, T3H, T2j, TC, TJ, T5A, T5B, T5C, T5D;
53 	       V T4G, T5h, T2d, T3F, T4L, T5g, T2m, T3I;
54 	       {
55 		    V T3, T1x, T14, T2R, T6, T2S, T17, T1y;
56 		    {
57 			 V T1, T2, T12, T13;
58 			 T1 = LD(&(ri[0]), ivs, &(ri[0]));
59 			 T2 = LD(&(ri[WS(is, 16)]), ivs, &(ri[0]));
60 			 T3 = VADD(T1, T2);
61 			 T1x = VSUB(T1, T2);
62 			 T12 = LD(&(ii[0]), ivs, &(ii[0]));
63 			 T13 = LD(&(ii[WS(is, 16)]), ivs, &(ii[0]));
64 			 T14 = VADD(T12, T13);
65 			 T2R = VSUB(T12, T13);
66 		    }
67 		    {
68 			 V T4, T5, T15, T16;
69 			 T4 = LD(&(ri[WS(is, 8)]), ivs, &(ri[0]));
70 			 T5 = LD(&(ri[WS(is, 24)]), ivs, &(ri[0]));
71 			 T6 = VADD(T4, T5);
72 			 T2S = VSUB(T4, T5);
73 			 T15 = LD(&(ii[WS(is, 8)]), ivs, &(ii[0]));
74 			 T16 = LD(&(ii[WS(is, 24)]), ivs, &(ii[0]));
75 			 T17 = VADD(T15, T16);
76 			 T1y = VSUB(T15, T16);
77 		    }
78 		    T7 = VADD(T3, T6);
79 		    T4r = VSUB(T3, T6);
80 		    T4Z = VSUB(T14, T17);
81 		    T18 = VADD(T14, T17);
82 		    T1z = VADD(T1x, T1y);
83 		    T3t = VSUB(T1x, T1y);
84 		    T3T = VADD(T2S, T2R);
85 		    T2T = VSUB(T2R, T2S);
86 	       }
87 	       {
88 		    V Ta, T1A, T1b, T1B, Td, T1D, T1e, T1E;
89 		    {
90 			 V T8, T9, T19, T1a;
91 			 T8 = LD(&(ri[WS(is, 4)]), ivs, &(ri[0]));
92 			 T9 = LD(&(ri[WS(is, 20)]), ivs, &(ri[0]));
93 			 Ta = VADD(T8, T9);
94 			 T1A = VSUB(T8, T9);
95 			 T19 = LD(&(ii[WS(is, 4)]), ivs, &(ii[0]));
96 			 T1a = LD(&(ii[WS(is, 20)]), ivs, &(ii[0]));
97 			 T1b = VADD(T19, T1a);
98 			 T1B = VSUB(T19, T1a);
99 		    }
100 		    {
101 			 V Tb, Tc, T1c, T1d;
102 			 Tb = LD(&(ri[WS(is, 28)]), ivs, &(ri[0]));
103 			 Tc = LD(&(ri[WS(is, 12)]), ivs, &(ri[0]));
104 			 Td = VADD(Tb, Tc);
105 			 T1D = VSUB(Tb, Tc);
106 			 T1c = LD(&(ii[WS(is, 28)]), ivs, &(ii[0]));
107 			 T1d = LD(&(ii[WS(is, 12)]), ivs, &(ii[0]));
108 			 T1e = VADD(T1c, T1d);
109 			 T1E = VSUB(T1c, T1d);
110 		    }
111 		    Te = VADD(Ta, Td);
112 		    T1f = VADD(T1b, T1e);
113 		    T50 = VSUB(Td, Ta);
114 		    T4s = VSUB(T1b, T1e);
115 		    {
116 			 V T2U, T2V, T1C, T1F;
117 			 T2U = VSUB(T1B, T1A);
118 			 T2V = VADD(T1D, T1E);
119 			 T2W = VADD(T2U, T2V);
120 			 T3u = VSUB(T2U, T2V);
121 			 T1C = VADD(T1A, T1B);
122 			 T1F = VSUB(T1D, T1E);
123 			 T1G = VADD(T1C, T1F);
124 			 T3U = VSUB(T1F, T1C);
125 		    }
126 	       }
127 	       {
128 		    V Ti, T1L, T1j, T1I, Tl, T1J, T1m, T1M, T1K, T1N;
129 		    {
130 			 V Tg, Th, T1h, T1i;
131 			 Tg = LD(&(ri[WS(is, 2)]), ivs, &(ri[0]));
132 			 Th = LD(&(ri[WS(is, 18)]), ivs, &(ri[0]));
133 			 Ti = VADD(Tg, Th);
134 			 T1L = VSUB(Tg, Th);
135 			 T1h = LD(&(ii[WS(is, 2)]), ivs, &(ii[0]));
136 			 T1i = LD(&(ii[WS(is, 18)]), ivs, &(ii[0]));
137 			 T1j = VADD(T1h, T1i);
138 			 T1I = VSUB(T1h, T1i);
139 		    }
140 		    {
141 			 V Tj, Tk, T1k, T1l;
142 			 Tj = LD(&(ri[WS(is, 10)]), ivs, &(ri[0]));
143 			 Tk = LD(&(ri[WS(is, 26)]), ivs, &(ri[0]));
144 			 Tl = VADD(Tj, Tk);
145 			 T1J = VSUB(Tj, Tk);
146 			 T1k = LD(&(ii[WS(is, 10)]), ivs, &(ii[0]));
147 			 T1l = LD(&(ii[WS(is, 26)]), ivs, &(ii[0]));
148 			 T1m = VADD(T1k, T1l);
149 			 T1M = VSUB(T1k, T1l);
150 		    }
151 		    Tm = VADD(Ti, Tl);
152 		    T1n = VADD(T1j, T1m);
153 		    T1K = VSUB(T1I, T1J);
154 		    T1N = VADD(T1L, T1M);
155 		    T1O = VFNMS(LDK(KP414213562), T1N, T1K);
156 		    T2Z = VFMA(LDK(KP414213562), T1K, T1N);
157 		    {
158 			 V T3w, T3x, T4u, T4v;
159 			 T3w = VADD(T1J, T1I);
160 			 T3x = VSUB(T1L, T1M);
161 			 T3y = VFMA(LDK(KP414213562), T3x, T3w);
162 			 T3X = VFNMS(LDK(KP414213562), T3w, T3x);
163 			 T4u = VSUB(T1j, T1m);
164 			 T4v = VSUB(Ti, Tl);
165 			 T4w = VSUB(T4u, T4v);
166 			 T53 = VADD(T4v, T4u);
167 		    }
168 	       }
169 	       {
170 		    V Tp, T1S, T1q, T1P, Ts, T1Q, T1t, T1T, T1R, T1U;
171 		    {
172 			 V Tn, To, T1o, T1p;
173 			 Tn = LD(&(ri[WS(is, 30)]), ivs, &(ri[0]));
174 			 To = LD(&(ri[WS(is, 14)]), ivs, &(ri[0]));
175 			 Tp = VADD(Tn, To);
176 			 T1S = VSUB(Tn, To);
177 			 T1o = LD(&(ii[WS(is, 30)]), ivs, &(ii[0]));
178 			 T1p = LD(&(ii[WS(is, 14)]), ivs, &(ii[0]));
179 			 T1q = VADD(T1o, T1p);
180 			 T1P = VSUB(T1o, T1p);
181 		    }
182 		    {
183 			 V Tq, Tr, T1r, T1s;
184 			 Tq = LD(&(ri[WS(is, 6)]), ivs, &(ri[0]));
185 			 Tr = LD(&(ri[WS(is, 22)]), ivs, &(ri[0]));
186 			 Ts = VADD(Tq, Tr);
187 			 T1Q = VSUB(Tq, Tr);
188 			 T1r = LD(&(ii[WS(is, 6)]), ivs, &(ii[0]));
189 			 T1s = LD(&(ii[WS(is, 22)]), ivs, &(ii[0]));
190 			 T1t = VADD(T1r, T1s);
191 			 T1T = VSUB(T1r, T1s);
192 		    }
193 		    Tt = VADD(Tp, Ts);
194 		    T1u = VADD(T1q, T1t);
195 		    T1R = VSUB(T1P, T1Q);
196 		    T1U = VADD(T1S, T1T);
197 		    T1V = VFMA(LDK(KP414213562), T1U, T1R);
198 		    T2Y = VFNMS(LDK(KP414213562), T1R, T1U);
199 		    {
200 			 V T3z, T3A, T4x, T4y;
201 			 T3z = VADD(T1Q, T1P);
202 			 T3A = VSUB(T1S, T1T);
203 			 T3B = VFNMS(LDK(KP414213562), T3A, T3z);
204 			 T3W = VFMA(LDK(KP414213562), T3z, T3A);
205 			 T4x = VSUB(Tp, Ts);
206 			 T4y = VSUB(T1q, T1t);
207 			 T4z = VADD(T4x, T4y);
208 			 T52 = VSUB(T4x, T4y);
209 		    }
210 	       }
211 	       {
212 		    V TN, T2G, T2r, T4N, TQ, T2s, T2J, T4O, TU, T2x, T2w, T4T, TX, T2z, T2C;
213 		    V T4U;
214 		    {
215 			 V TL, TM, T2p, T2q;
216 			 TL = LD(&(ri[WS(is, 31)]), ivs, &(ri[WS(is, 1)]));
217 			 TM = LD(&(ri[WS(is, 15)]), ivs, &(ri[WS(is, 1)]));
218 			 TN = VADD(TL, TM);
219 			 T2G = VSUB(TL, TM);
220 			 T2p = LD(&(ii[WS(is, 31)]), ivs, &(ii[WS(is, 1)]));
221 			 T2q = LD(&(ii[WS(is, 15)]), ivs, &(ii[WS(is, 1)]));
222 			 T2r = VSUB(T2p, T2q);
223 			 T4N = VADD(T2p, T2q);
224 		    }
225 		    {
226 			 V TO, TP, T2H, T2I;
227 			 TO = LD(&(ri[WS(is, 7)]), ivs, &(ri[WS(is, 1)]));
228 			 TP = LD(&(ri[WS(is, 23)]), ivs, &(ri[WS(is, 1)]));
229 			 TQ = VADD(TO, TP);
230 			 T2s = VSUB(TO, TP);
231 			 T2H = LD(&(ii[WS(is, 7)]), ivs, &(ii[WS(is, 1)]));
232 			 T2I = LD(&(ii[WS(is, 23)]), ivs, &(ii[WS(is, 1)]));
233 			 T2J = VSUB(T2H, T2I);
234 			 T4O = VADD(T2H, T2I);
235 		    }
236 		    {
237 			 V TS, TT, T2u, T2v;
238 			 TS = LD(&(ri[WS(is, 3)]), ivs, &(ri[WS(is, 1)]));
239 			 TT = LD(&(ri[WS(is, 19)]), ivs, &(ri[WS(is, 1)]));
240 			 TU = VADD(TS, TT);
241 			 T2x = VSUB(TS, TT);
242 			 T2u = LD(&(ii[WS(is, 3)]), ivs, &(ii[WS(is, 1)]));
243 			 T2v = LD(&(ii[WS(is, 19)]), ivs, &(ii[WS(is, 1)]));
244 			 T2w = VSUB(T2u, T2v);
245 			 T4T = VADD(T2u, T2v);
246 		    }
247 		    {
248 			 V TV, TW, T2A, T2B;
249 			 TV = LD(&(ri[WS(is, 27)]), ivs, &(ri[WS(is, 1)]));
250 			 TW = LD(&(ri[WS(is, 11)]), ivs, &(ri[WS(is, 1)]));
251 			 TX = VADD(TV, TW);
252 			 T2z = VSUB(TV, TW);
253 			 T2A = LD(&(ii[WS(is, 27)]), ivs, &(ii[WS(is, 1)]));
254 			 T2B = LD(&(ii[WS(is, 11)]), ivs, &(ii[WS(is, 1)]));
255 			 T2C = VSUB(T2A, T2B);
256 			 T4U = VADD(T2A, T2B);
257 		    }
258 		    T2t = VSUB(T2r, T2s);
259 		    T3L = VSUB(T2G, T2J);
260 		    T3O = VADD(T2s, T2r);
261 		    T2K = VADD(T2G, T2J);
262 		    TR = VADD(TN, TQ);
263 		    TY = VADD(TU, TX);
264 		    T5F = VSUB(TR, TY);
265 		    {
266 			 V T4P, T4Q, T2y, T2D;
267 			 T5G = VADD(T4N, T4O);
268 			 T5H = VADD(T4T, T4U);
269 			 T5I = VSUB(T5G, T5H);
270 			 T4P = VSUB(T4N, T4O);
271 			 T4Q = VSUB(TX, TU);
272 			 T4R = VSUB(T4P, T4Q);
273 			 T5k = VADD(T4Q, T4P);
274 			 T2y = VSUB(T2w, T2x);
275 			 T2D = VADD(T2z, T2C);
276 			 T2E = VADD(T2y, T2D);
277 			 T3M = VSUB(T2D, T2y);
278 			 {
279 			      V T4S, T4V, T2L, T2M;
280 			      T4S = VSUB(TN, TQ);
281 			      T4V = VSUB(T4T, T4U);
282 			      T4W = VSUB(T4S, T4V);
283 			      T5j = VADD(T4S, T4V);
284 			      T2L = VADD(T2x, T2w);
285 			      T2M = VSUB(T2z, T2C);
286 			      T2N = VADD(T2L, T2M);
287 			      T3P = VSUB(T2L, T2M);
288 			 }
289 		    }
290 	       }
291 	       {
292 		    V Ty, T2f, T20, T4C, TB, T21, T2i, T4D, TF, T26, T25, T4I, TI, T28, T2b;
293 		    V T4J;
294 		    {
295 			 V Tw, Tx, T1Y, T1Z;
296 			 Tw = LD(&(ri[WS(is, 1)]), ivs, &(ri[WS(is, 1)]));
297 			 Tx = LD(&(ri[WS(is, 17)]), ivs, &(ri[WS(is, 1)]));
298 			 Ty = VADD(Tw, Tx);
299 			 T2f = VSUB(Tw, Tx);
300 			 T1Y = LD(&(ii[WS(is, 1)]), ivs, &(ii[WS(is, 1)]));
301 			 T1Z = LD(&(ii[WS(is, 17)]), ivs, &(ii[WS(is, 1)]));
302 			 T20 = VSUB(T1Y, T1Z);
303 			 T4C = VADD(T1Y, T1Z);
304 		    }
305 		    {
306 			 V Tz, TA, T2g, T2h;
307 			 Tz = LD(&(ri[WS(is, 9)]), ivs, &(ri[WS(is, 1)]));
308 			 TA = LD(&(ri[WS(is, 25)]), ivs, &(ri[WS(is, 1)]));
309 			 TB = VADD(Tz, TA);
310 			 T21 = VSUB(Tz, TA);
311 			 T2g = LD(&(ii[WS(is, 9)]), ivs, &(ii[WS(is, 1)]));
312 			 T2h = LD(&(ii[WS(is, 25)]), ivs, &(ii[WS(is, 1)]));
313 			 T2i = VSUB(T2g, T2h);
314 			 T4D = VADD(T2g, T2h);
315 		    }
316 		    {
317 			 V TD, TE, T23, T24;
318 			 TD = LD(&(ri[WS(is, 5)]), ivs, &(ri[WS(is, 1)]));
319 			 TE = LD(&(ri[WS(is, 21)]), ivs, &(ri[WS(is, 1)]));
320 			 TF = VADD(TD, TE);
321 			 T26 = VSUB(TD, TE);
322 			 T23 = LD(&(ii[WS(is, 5)]), ivs, &(ii[WS(is, 1)]));
323 			 T24 = LD(&(ii[WS(is, 21)]), ivs, &(ii[WS(is, 1)]));
324 			 T25 = VSUB(T23, T24);
325 			 T4I = VADD(T23, T24);
326 		    }
327 		    {
328 			 V TG, TH, T29, T2a;
329 			 TG = LD(&(ri[WS(is, 29)]), ivs, &(ri[WS(is, 1)]));
330 			 TH = LD(&(ri[WS(is, 13)]), ivs, &(ri[WS(is, 1)]));
331 			 TI = VADD(TG, TH);
332 			 T28 = VSUB(TG, TH);
333 			 T29 = LD(&(ii[WS(is, 29)]), ivs, &(ii[WS(is, 1)]));
334 			 T2a = LD(&(ii[WS(is, 13)]), ivs, &(ii[WS(is, 1)]));
335 			 T2b = VSUB(T29, T2a);
336 			 T4J = VADD(T29, T2a);
337 		    }
338 		    T22 = VSUB(T20, T21);
339 		    T3E = VSUB(T2f, T2i);
340 		    T3H = VADD(T21, T20);
341 		    T2j = VADD(T2f, T2i);
342 		    TC = VADD(Ty, TB);
343 		    TJ = VADD(TF, TI);
344 		    T5A = VSUB(TC, TJ);
345 		    {
346 			 V T4E, T4F, T27, T2c;
347 			 T5B = VADD(T4C, T4D);
348 			 T5C = VADD(T4I, T4J);
349 			 T5D = VSUB(T5B, T5C);
350 			 T4E = VSUB(T4C, T4D);
351 			 T4F = VSUB(TI, TF);
352 			 T4G = VSUB(T4E, T4F);
353 			 T5h = VADD(T4F, T4E);
354 			 T27 = VSUB(T25, T26);
355 			 T2c = VADD(T28, T2b);
356 			 T2d = VADD(T27, T2c);
357 			 T3F = VSUB(T2c, T27);
358 			 {
359 			      V T4H, T4K, T2k, T2l;
360 			      T4H = VSUB(Ty, TB);
361 			      T4K = VSUB(T4I, T4J);
362 			      T4L = VSUB(T4H, T4K);
363 			      T5g = VADD(T4H, T4K);
364 			      T2k = VADD(T26, T25);
365 			      T2l = VSUB(T28, T2b);
366 			      T2m = VADD(T2k, T2l);
367 			      T3I = VSUB(T2k, T2l);
368 			 }
369 		    }
370 	       }
371 	       {
372 		    V T61, T62, T63, T64, T65, T66, T67, T68, T69, T6a, T6b, T6c, T6d, T6e, T6f;
373 		    V T6g, T6h, T6i, T6j, T6k, T6l, T6m, T6n, T6o, T6p, T6q, T6r, T6s, T6t, T6u;
374 		    V T6v, T6w;
375 		    {
376 			 V T4B, T5b, T5a, T5c, T4Y, T56, T55, T57;
377 			 {
378 			      V T4t, T4A, T58, T59;
379 			      T4t = VSUB(T4r, T4s);
380 			      T4A = VSUB(T4w, T4z);
381 			      T4B = VFMA(LDK(KP707106781), T4A, T4t);
382 			      T5b = VFNMS(LDK(KP707106781), T4A, T4t);
383 			      T58 = VFMA(LDK(KP414213562), T4R, T4W);
384 			      T59 = VFNMS(LDK(KP414213562), T4G, T4L);
385 			      T5a = VSUB(T58, T59);
386 			      T5c = VADD(T59, T58);
387 			 }
388 			 {
389 			      V T4M, T4X, T51, T54;
390 			      T4M = VFMA(LDK(KP414213562), T4L, T4G);
391 			      T4X = VFNMS(LDK(KP414213562), T4W, T4R);
392 			      T4Y = VSUB(T4M, T4X);
393 			      T56 = VADD(T4M, T4X);
394 			      T51 = VSUB(T4Z, T50);
395 			      T54 = VSUB(T52, T53);
396 			      T55 = VFNMS(LDK(KP707106781), T54, T51);
397 			      T57 = VFMA(LDK(KP707106781), T54, T51);
398 			 }
399 			 T61 = VFNMS(LDK(KP923879532), T4Y, T4B);
400 			 STM4(&(ro[22]), T61, ovs, &(ro[0]));
401 			 T62 = VFNMS(LDK(KP923879532), T5a, T57);
402 			 STM4(&(io[22]), T62, ovs, &(io[0]));
403 			 T63 = VFMA(LDK(KP923879532), T4Y, T4B);
404 			 STM4(&(ro[6]), T63, ovs, &(ro[0]));
405 			 T64 = VFMA(LDK(KP923879532), T5a, T57);
406 			 STM4(&(io[6]), T64, ovs, &(io[0]));
407 			 T65 = VFNMS(LDK(KP923879532), T56, T55);
408 			 STM4(&(io[14]), T65, ovs, &(io[0]));
409 			 T66 = VFNMS(LDK(KP923879532), T5c, T5b);
410 			 STM4(&(ro[14]), T66, ovs, &(ro[0]));
411 			 T67 = VFMA(LDK(KP923879532), T56, T55);
412 			 STM4(&(io[30]), T67, ovs, &(io[0]));
413 			 T68 = VFMA(LDK(KP923879532), T5c, T5b);
414 			 STM4(&(ro[30]), T68, ovs, &(ro[0]));
415 		    }
416 		    {
417 			 V T5f, T5r, T5u, T5w, T5m, T5q, T5p, T5v;
418 			 {
419 			      V T5d, T5e, T5s, T5t;
420 			      T5d = VADD(T4r, T4s);
421 			      T5e = VADD(T53, T52);
422 			      T5f = VFMA(LDK(KP707106781), T5e, T5d);
423 			      T5r = VFNMS(LDK(KP707106781), T5e, T5d);
424 			      T5s = VFNMS(LDK(KP414213562), T5g, T5h);
425 			      T5t = VFMA(LDK(KP414213562), T5j, T5k);
426 			      T5u = VSUB(T5s, T5t);
427 			      T5w = VADD(T5s, T5t);
428 			 }
429 			 {
430 			      V T5i, T5l, T5n, T5o;
431 			      T5i = VFMA(LDK(KP414213562), T5h, T5g);
432 			      T5l = VFNMS(LDK(KP414213562), T5k, T5j);
433 			      T5m = VADD(T5i, T5l);
434 			      T5q = VSUB(T5l, T5i);
435 			      T5n = VADD(T50, T4Z);
436 			      T5o = VADD(T4w, T4z);
437 			      T5p = VFNMS(LDK(KP707106781), T5o, T5n);
438 			      T5v = VFMA(LDK(KP707106781), T5o, T5n);
439 			 }
440 			 T69 = VFNMS(LDK(KP923879532), T5m, T5f);
441 			 STM4(&(ro[18]), T69, ovs, &(ro[0]));
442 			 T6a = VFNMS(LDK(KP923879532), T5w, T5v);
443 			 STM4(&(io[18]), T6a, ovs, &(io[0]));
444 			 T6b = VFMA(LDK(KP923879532), T5m, T5f);
445 			 STM4(&(ro[2]), T6b, ovs, &(ro[0]));
446 			 T6c = VFMA(LDK(KP923879532), T5w, T5v);
447 			 STM4(&(io[2]), T6c, ovs, &(io[0]));
448 			 T6d = VFNMS(LDK(KP923879532), T5q, T5p);
449 			 STM4(&(io[26]), T6d, ovs, &(io[0]));
450 			 T6e = VFNMS(LDK(KP923879532), T5u, T5r);
451 			 STM4(&(ro[26]), T6e, ovs, &(ro[0]));
452 			 T6f = VFMA(LDK(KP923879532), T5q, T5p);
453 			 STM4(&(io[10]), T6f, ovs, &(io[0]));
454 			 T6g = VFMA(LDK(KP923879532), T5u, T5r);
455 			 STM4(&(ro[10]), T6g, ovs, &(ro[0]));
456 		    }
457 		    {
458 			 V T5z, T5P, T5S, T5U, T5K, T5O, T5N, T5T;
459 			 {
460 			      V T5x, T5y, T5Q, T5R;
461 			      T5x = VSUB(T7, Te);
462 			      T5y = VSUB(T1n, T1u);
463 			      T5z = VADD(T5x, T5y);
464 			      T5P = VSUB(T5x, T5y);
465 			      T5Q = VSUB(T5D, T5A);
466 			      T5R = VADD(T5F, T5I);
467 			      T5S = VSUB(T5Q, T5R);
468 			      T5U = VADD(T5Q, T5R);
469 			 }
470 			 {
471 			      V T5E, T5J, T5L, T5M;
472 			      T5E = VADD(T5A, T5D);
473 			      T5J = VSUB(T5F, T5I);
474 			      T5K = VADD(T5E, T5J);
475 			      T5O = VSUB(T5J, T5E);
476 			      T5L = VSUB(T18, T1f);
477 			      T5M = VSUB(Tt, Tm);
478 			      T5N = VSUB(T5L, T5M);
479 			      T5T = VADD(T5M, T5L);
480 			 }
481 			 T6h = VFNMS(LDK(KP707106781), T5K, T5z);
482 			 STM4(&(ro[20]), T6h, ovs, &(ro[0]));
483 			 T6i = VFNMS(LDK(KP707106781), T5U, T5T);
484 			 STM4(&(io[20]), T6i, ovs, &(io[0]));
485 			 T6j = VFMA(LDK(KP707106781), T5K, T5z);
486 			 STM4(&(ro[4]), T6j, ovs, &(ro[0]));
487 			 T6k = VFMA(LDK(KP707106781), T5U, T5T);
488 			 STM4(&(io[4]), T6k, ovs, &(io[0]));
489 			 T6l = VFNMS(LDK(KP707106781), T5O, T5N);
490 			 STM4(&(io[28]), T6l, ovs, &(io[0]));
491 			 T6m = VFNMS(LDK(KP707106781), T5S, T5P);
492 			 STM4(&(ro[28]), T6m, ovs, &(ro[0]));
493 			 T6n = VFMA(LDK(KP707106781), T5O, T5N);
494 			 STM4(&(io[12]), T6n, ovs, &(io[0]));
495 			 T6o = VFMA(LDK(KP707106781), T5S, T5P);
496 			 STM4(&(ro[12]), T6o, ovs, &(ro[0]));
497 		    }
498 		    {
499 			 V Tv, T5V, T5Y, T60, T10, T11, T1w, T5Z;
500 			 {
501 			      V Tf, Tu, T5W, T5X;
502 			      Tf = VADD(T7, Te);
503 			      Tu = VADD(Tm, Tt);
504 			      Tv = VADD(Tf, Tu);
505 			      T5V = VSUB(Tf, Tu);
506 			      T5W = VADD(T5B, T5C);
507 			      T5X = VADD(T5G, T5H);
508 			      T5Y = VSUB(T5W, T5X);
509 			      T60 = VADD(T5W, T5X);
510 			 }
511 			 {
512 			      V TK, TZ, T1g, T1v;
513 			      TK = VADD(TC, TJ);
514 			      TZ = VADD(TR, TY);
515 			      T10 = VADD(TK, TZ);
516 			      T11 = VSUB(TZ, TK);
517 			      T1g = VADD(T18, T1f);
518 			      T1v = VADD(T1n, T1u);
519 			      T1w = VSUB(T1g, T1v);
520 			      T5Z = VADD(T1g, T1v);
521 			 }
522 			 T6p = VSUB(Tv, T10);
523 			 STM4(&(ro[16]), T6p, ovs, &(ro[0]));
524 			 T6q = VSUB(T5Z, T60);
525 			 STM4(&(io[16]), T6q, ovs, &(io[0]));
526 			 T6r = VADD(Tv, T10);
527 			 STM4(&(ro[0]), T6r, ovs, &(ro[0]));
528 			 T6s = VADD(T5Z, T60);
529 			 STM4(&(io[0]), T6s, ovs, &(io[0]));
530 			 T6t = VADD(T11, T1w);
531 			 STM4(&(io[8]), T6t, ovs, &(io[0]));
532 			 T6u = VADD(T5V, T5Y);
533 			 STM4(&(ro[8]), T6u, ovs, &(ro[0]));
534 			 T6v = VSUB(T1w, T11);
535 			 STM4(&(io[24]), T6v, ovs, &(io[0]));
536 			 T6w = VSUB(T5V, T5Y);
537 			 STM4(&(ro[24]), T6w, ovs, &(ro[0]));
538 		    }
539 		    {
540 			 V T6x, T6y, T6z, T6A, T6B, T6C, T6D, T6E, T6F, T6G, T6H, T6I, T6J, T6K, T6L;
541 			 V T6M;
542 			 {
543 			      V T1X, T37, T31, T33, T2o, T35, T2P, T34;
544 			      {
545 				   V T1H, T1W, T2X, T30;
546 				   T1H = VFNMS(LDK(KP707106781), T1G, T1z);
547 				   T1W = VSUB(T1O, T1V);
548 				   T1X = VFMA(LDK(KP923879532), T1W, T1H);
549 				   T37 = VFNMS(LDK(KP923879532), T1W, T1H);
550 				   T2X = VFNMS(LDK(KP707106781), T2W, T2T);
551 				   T30 = VSUB(T2Y, T2Z);
552 				   T31 = VFNMS(LDK(KP923879532), T30, T2X);
553 				   T33 = VFMA(LDK(KP923879532), T30, T2X);
554 			      }
555 			      {
556 				   V T2e, T2n, T2F, T2O;
557 				   T2e = VFNMS(LDK(KP707106781), T2d, T22);
558 				   T2n = VFNMS(LDK(KP707106781), T2m, T2j);
559 				   T2o = VFMA(LDK(KP668178637), T2n, T2e);
560 				   T35 = VFNMS(LDK(KP668178637), T2e, T2n);
561 				   T2F = VFNMS(LDK(KP707106781), T2E, T2t);
562 				   T2O = VFNMS(LDK(KP707106781), T2N, T2K);
563 				   T2P = VFNMS(LDK(KP668178637), T2O, T2F);
564 				   T34 = VFMA(LDK(KP668178637), T2F, T2O);
565 			      }
566 			      {
567 				   V T2Q, T36, T32, T38;
568 				   T2Q = VSUB(T2o, T2P);
569 				   T6x = VFNMS(LDK(KP831469612), T2Q, T1X);
570 				   STM4(&(ro[21]), T6x, ovs, &(ro[1]));
571 				   T6y = VFMA(LDK(KP831469612), T2Q, T1X);
572 				   STM4(&(ro[5]), T6y, ovs, &(ro[1]));
573 				   T36 = VSUB(T34, T35);
574 				   T6z = VFNMS(LDK(KP831469612), T36, T33);
575 				   STM4(&(io[21]), T6z, ovs, &(io[1]));
576 				   T6A = VFMA(LDK(KP831469612), T36, T33);
577 				   STM4(&(io[5]), T6A, ovs, &(io[1]));
578 				   T32 = VADD(T2o, T2P);
579 				   T6B = VFNMS(LDK(KP831469612), T32, T31);
580 				   STM4(&(io[13]), T6B, ovs, &(io[1]));
581 				   T6C = VFMA(LDK(KP831469612), T32, T31);
582 				   STM4(&(io[29]), T6C, ovs, &(io[1]));
583 				   T38 = VADD(T35, T34);
584 				   T6D = VFNMS(LDK(KP831469612), T38, T37);
585 				   STM4(&(ro[13]), T6D, ovs, &(ro[1]));
586 				   T6E = VFMA(LDK(KP831469612), T38, T37);
587 				   STM4(&(ro[29]), T6E, ovs, &(ro[1]));
588 			      }
589 			 }
590 			 {
591 			      V T3D, T41, T3Z, T45, T3K, T42, T3R, T43;
592 			      {
593 				   V T3v, T3C, T3V, T3Y;
594 				   T3v = VFMA(LDK(KP707106781), T3u, T3t);
595 				   T3C = VSUB(T3y, T3B);
596 				   T3D = VFMA(LDK(KP923879532), T3C, T3v);
597 				   T41 = VFNMS(LDK(KP923879532), T3C, T3v);
598 				   T3V = VFMA(LDK(KP707106781), T3U, T3T);
599 				   T3Y = VSUB(T3W, T3X);
600 				   T3Z = VFNMS(LDK(KP923879532), T3Y, T3V);
601 				   T45 = VFMA(LDK(KP923879532), T3Y, T3V);
602 			      }
603 			      {
604 				   V T3G, T3J, T3N, T3Q;
605 				   T3G = VFNMS(LDK(KP707106781), T3F, T3E);
606 				   T3J = VFNMS(LDK(KP707106781), T3I, T3H);
607 				   T3K = VFMA(LDK(KP668178637), T3J, T3G);
608 				   T42 = VFNMS(LDK(KP668178637), T3G, T3J);
609 				   T3N = VFNMS(LDK(KP707106781), T3M, T3L);
610 				   T3Q = VFNMS(LDK(KP707106781), T3P, T3O);
611 				   T3R = VFNMS(LDK(KP668178637), T3Q, T3N);
612 				   T43 = VFMA(LDK(KP668178637), T3N, T3Q);
613 			      }
614 			      {
615 				   V T3S, T46, T40, T44;
616 				   T3S = VADD(T3K, T3R);
617 				   T6F = VFNMS(LDK(KP831469612), T3S, T3D);
618 				   STM4(&(ro[19]), T6F, ovs, &(ro[1]));
619 				   T6G = VFMA(LDK(KP831469612), T3S, T3D);
620 				   STM4(&(ro[3]), T6G, ovs, &(ro[1]));
621 				   T46 = VADD(T42, T43);
622 				   T6H = VFNMS(LDK(KP831469612), T46, T45);
623 				   STM4(&(io[19]), T6H, ovs, &(io[1]));
624 				   T6I = VFMA(LDK(KP831469612), T46, T45);
625 				   STM4(&(io[3]), T6I, ovs, &(io[1]));
626 				   T40 = VSUB(T3R, T3K);
627 				   T6J = VFNMS(LDK(KP831469612), T40, T3Z);
628 				   STM4(&(io[27]), T6J, ovs, &(io[1]));
629 				   T6K = VFMA(LDK(KP831469612), T40, T3Z);
630 				   STM4(&(io[11]), T6K, ovs, &(io[1]));
631 				   T44 = VSUB(T42, T43);
632 				   T6L = VFNMS(LDK(KP831469612), T44, T41);
633 				   STM4(&(ro[27]), T6L, ovs, &(ro[1]));
634 				   T6M = VFMA(LDK(KP831469612), T44, T41);
635 				   STM4(&(ro[11]), T6M, ovs, &(ro[1]));
636 			      }
637 			 }
638 			 {
639 			      V T49, T4p, T4j, T4l, T4c, T4n, T4f, T4m;
640 			      {
641 				   V T47, T48, T4h, T4i;
642 				   T47 = VFNMS(LDK(KP707106781), T3u, T3t);
643 				   T48 = VADD(T3X, T3W);
644 				   T49 = VFNMS(LDK(KP923879532), T48, T47);
645 				   T4p = VFMA(LDK(KP923879532), T48, T47);
646 				   T4h = VFNMS(LDK(KP707106781), T3U, T3T);
647 				   T4i = VADD(T3y, T3B);
648 				   T4j = VFMA(LDK(KP923879532), T4i, T4h);
649 				   T4l = VFNMS(LDK(KP923879532), T4i, T4h);
650 			      }
651 			      {
652 				   V T4a, T4b, T4d, T4e;
653 				   T4a = VFMA(LDK(KP707106781), T3I, T3H);
654 				   T4b = VFMA(LDK(KP707106781), T3F, T3E);
655 				   T4c = VFMA(LDK(KP198912367), T4b, T4a);
656 				   T4n = VFNMS(LDK(KP198912367), T4a, T4b);
657 				   T4d = VFMA(LDK(KP707106781), T3P, T3O);
658 				   T4e = VFMA(LDK(KP707106781), T3M, T3L);
659 				   T4f = VFNMS(LDK(KP198912367), T4e, T4d);
660 				   T4m = VFMA(LDK(KP198912367), T4d, T4e);
661 			      }
662 			      {
663 				   V T4g, T6N, T6O, T4o, T6P, T6Q;
664 				   T4g = VSUB(T4c, T4f);
665 				   T6N = VFNMS(LDK(KP980785280), T4g, T49);
666 				   STM4(&(ro[23]), T6N, ovs, &(ro[1]));
667 				   STN4(&(ro[20]), T6h, T6x, T61, T6N, ovs);
668 				   T6O = VFMA(LDK(KP980785280), T4g, T49);
669 				   STM4(&(ro[7]), T6O, ovs, &(ro[1]));
670 				   STN4(&(ro[4]), T6j, T6y, T63, T6O, ovs);
671 				   T4o = VSUB(T4m, T4n);
672 				   T6P = VFNMS(LDK(KP980785280), T4o, T4l);
673 				   STM4(&(io[23]), T6P, ovs, &(io[1]));
674 				   STN4(&(io[20]), T6i, T6z, T62, T6P, ovs);
675 				   T6Q = VFMA(LDK(KP980785280), T4o, T4l);
676 				   STM4(&(io[7]), T6Q, ovs, &(io[1]));
677 				   STN4(&(io[4]), T6k, T6A, T64, T6Q, ovs);
678 			      }
679 			      {
680 				   V T4k, T6R, T6S, T4q, T6T, T6U;
681 				   T4k = VADD(T4c, T4f);
682 				   T6R = VFNMS(LDK(KP980785280), T4k, T4j);
683 				   STM4(&(io[15]), T6R, ovs, &(io[1]));
684 				   STN4(&(io[12]), T6n, T6B, T65, T6R, ovs);
685 				   T6S = VFMA(LDK(KP980785280), T4k, T4j);
686 				   STM4(&(io[31]), T6S, ovs, &(io[1]));
687 				   STN4(&(io[28]), T6l, T6C, T67, T6S, ovs);
688 				   T4q = VADD(T4n, T4m);
689 				   T6T = VFNMS(LDK(KP980785280), T4q, T4p);
690 				   STM4(&(ro[15]), T6T, ovs, &(ro[1]));
691 				   STN4(&(ro[12]), T6o, T6D, T66, T6T, ovs);
692 				   T6U = VFMA(LDK(KP980785280), T4q, T4p);
693 				   STM4(&(ro[31]), T6U, ovs, &(ro[1]));
694 				   STN4(&(ro[28]), T6m, T6E, T68, T6U, ovs);
695 			      }
696 			 }
697 			 {
698 			      V T3b, T3n, T3l, T3r, T3e, T3o, T3h, T3p;
699 			      {
700 				   V T39, T3a, T3j, T3k;
701 				   T39 = VFMA(LDK(KP707106781), T1G, T1z);
702 				   T3a = VADD(T2Z, T2Y);
703 				   T3b = VFMA(LDK(KP923879532), T3a, T39);
704 				   T3n = VFNMS(LDK(KP923879532), T3a, T39);
705 				   T3j = VFMA(LDK(KP707106781), T2W, T2T);
706 				   T3k = VADD(T1O, T1V);
707 				   T3l = VFNMS(LDK(KP923879532), T3k, T3j);
708 				   T3r = VFMA(LDK(KP923879532), T3k, T3j);
709 			      }
710 			      {
711 				   V T3c, T3d, T3f, T3g;
712 				   T3c = VFMA(LDK(KP707106781), T2m, T2j);
713 				   T3d = VFMA(LDK(KP707106781), T2d, T22);
714 				   T3e = VFMA(LDK(KP198912367), T3d, T3c);
715 				   T3o = VFNMS(LDK(KP198912367), T3c, T3d);
716 				   T3f = VFMA(LDK(KP707106781), T2N, T2K);
717 				   T3g = VFMA(LDK(KP707106781), T2E, T2t);
718 				   T3h = VFNMS(LDK(KP198912367), T3g, T3f);
719 				   T3p = VFMA(LDK(KP198912367), T3f, T3g);
720 			      }
721 			      {
722 				   V T3i, T6V, T6W, T3s, T6X, T6Y;
723 				   T3i = VADD(T3e, T3h);
724 				   T6V = VFNMS(LDK(KP980785280), T3i, T3b);
725 				   STM4(&(ro[17]), T6V, ovs, &(ro[1]));
726 				   STN4(&(ro[16]), T6p, T6V, T69, T6F, ovs);
727 				   T6W = VFMA(LDK(KP980785280), T3i, T3b);
728 				   STM4(&(ro[1]), T6W, ovs, &(ro[1]));
729 				   STN4(&(ro[0]), T6r, T6W, T6b, T6G, ovs);
730 				   T3s = VADD(T3o, T3p);
731 				   T6X = VFNMS(LDK(KP980785280), T3s, T3r);
732 				   STM4(&(io[17]), T6X, ovs, &(io[1]));
733 				   STN4(&(io[16]), T6q, T6X, T6a, T6H, ovs);
734 				   T6Y = VFMA(LDK(KP980785280), T3s, T3r);
735 				   STM4(&(io[1]), T6Y, ovs, &(io[1]));
736 				   STN4(&(io[0]), T6s, T6Y, T6c, T6I, ovs);
737 			      }
738 			      {
739 				   V T3m, T6Z, T70, T3q, T71, T72;
740 				   T3m = VSUB(T3h, T3e);
741 				   T6Z = VFNMS(LDK(KP980785280), T3m, T3l);
742 				   STM4(&(io[25]), T6Z, ovs, &(io[1]));
743 				   STN4(&(io[24]), T6v, T6Z, T6d, T6J, ovs);
744 				   T70 = VFMA(LDK(KP980785280), T3m, T3l);
745 				   STM4(&(io[9]), T70, ovs, &(io[1]));
746 				   STN4(&(io[8]), T6t, T70, T6f, T6K, ovs);
747 				   T3q = VSUB(T3o, T3p);
748 				   T71 = VFNMS(LDK(KP980785280), T3q, T3n);
749 				   STM4(&(ro[25]), T71, ovs, &(ro[1]));
750 				   STN4(&(ro[24]), T6w, T71, T6e, T6L, ovs);
751 				   T72 = VFMA(LDK(KP980785280), T3q, T3n);
752 				   STM4(&(ro[9]), T72, ovs, &(ro[1]));
753 				   STN4(&(ro[8]), T6u, T72, T6g, T6M, ovs);
754 			      }
755 			 }
756 		    }
757 	       }
758 	  }
759      }
760      VLEAVE();
761 }
762 
763 static const kdft_desc desc = { 32, XSIMD_STRING("n2sv_32"), { 236, 0, 136, 0 }, &GENUS, 0, 1, 0, 0 };
764 
XSIMD(codelet_n2sv_32)765 void XSIMD(codelet_n2sv_32) (planner *p) { X(kdft_register) (p, n2sv_32, &desc);
766 }
767 
768 #else
769 
770 /* Generated by: ../../../genfft/gen_notw.native -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n2sv_32 -with-ostride 1 -include dft/simd/n2s.h -store-multiple 4 */
771 
772 /*
773  * This function contains 372 FP additions, 84 FP multiplications,
774  * (or, 340 additions, 52 multiplications, 32 fused multiply/add),
775  * 130 stack variables, 7 constants, and 144 memory accesses
776  */
777 #include "dft/simd/n2s.h"
778 
n2sv_32(const R * ri,const R * ii,R * ro,R * io,stride is,stride os,INT v,INT ivs,INT ovs)779 static void n2sv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
780 {
781      DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
782      DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
783      DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
784      DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
785      DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
786      DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
787      DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
788      {
789 	  INT i;
790 	  for (i = v; i > 0; i = i - (2 * VL), ri = ri + ((2 * VL) * ivs), ii = ii + ((2 * VL) * ivs), ro = ro + ((2 * VL) * ovs), io = io + ((2 * VL) * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) {
791 	       V T7, T4r, T4Z, T18, T1z, T3t, T3T, T2T, Te, T1f, T50, T4s, T2W, T3u, T1G;
792 	       V T3U, Tm, T1n, T1O, T2Z, T3y, T3X, T4w, T53, Tt, T1u, T1V, T2Y, T3B, T3W;
793 	       V T4z, T52, T2t, T3L, T3O, T2K, TR, TY, T5F, T5G, T5H, T5I, T4R, T5j, T2E;
794 	       V T3P, T4W, T5k, T2N, T3M, T22, T3E, T3H, T2j, TC, TJ, T5A, T5B, T5C, T5D;
795 	       V T4G, T5g, T2d, T3F, T4L, T5h, T2m, T3I;
796 	       {
797 		    V T3, T1x, T14, T2S, T6, T2R, T17, T1y;
798 		    {
799 			 V T1, T2, T12, T13;
800 			 T1 = LD(&(ri[0]), ivs, &(ri[0]));
801 			 T2 = LD(&(ri[WS(is, 16)]), ivs, &(ri[0]));
802 			 T3 = VADD(T1, T2);
803 			 T1x = VSUB(T1, T2);
804 			 T12 = LD(&(ii[0]), ivs, &(ii[0]));
805 			 T13 = LD(&(ii[WS(is, 16)]), ivs, &(ii[0]));
806 			 T14 = VADD(T12, T13);
807 			 T2S = VSUB(T12, T13);
808 		    }
809 		    {
810 			 V T4, T5, T15, T16;
811 			 T4 = LD(&(ri[WS(is, 8)]), ivs, &(ri[0]));
812 			 T5 = LD(&(ri[WS(is, 24)]), ivs, &(ri[0]));
813 			 T6 = VADD(T4, T5);
814 			 T2R = VSUB(T4, T5);
815 			 T15 = LD(&(ii[WS(is, 8)]), ivs, &(ii[0]));
816 			 T16 = LD(&(ii[WS(is, 24)]), ivs, &(ii[0]));
817 			 T17 = VADD(T15, T16);
818 			 T1y = VSUB(T15, T16);
819 		    }
820 		    T7 = VADD(T3, T6);
821 		    T4r = VSUB(T3, T6);
822 		    T4Z = VSUB(T14, T17);
823 		    T18 = VADD(T14, T17);
824 		    T1z = VSUB(T1x, T1y);
825 		    T3t = VADD(T1x, T1y);
826 		    T3T = VSUB(T2S, T2R);
827 		    T2T = VADD(T2R, T2S);
828 	       }
829 	       {
830 		    V Ta, T1B, T1b, T1A, Td, T1D, T1e, T1E;
831 		    {
832 			 V T8, T9, T19, T1a;
833 			 T8 = LD(&(ri[WS(is, 4)]), ivs, &(ri[0]));
834 			 T9 = LD(&(ri[WS(is, 20)]), ivs, &(ri[0]));
835 			 Ta = VADD(T8, T9);
836 			 T1B = VSUB(T8, T9);
837 			 T19 = LD(&(ii[WS(is, 4)]), ivs, &(ii[0]));
838 			 T1a = LD(&(ii[WS(is, 20)]), ivs, &(ii[0]));
839 			 T1b = VADD(T19, T1a);
840 			 T1A = VSUB(T19, T1a);
841 		    }
842 		    {
843 			 V Tb, Tc, T1c, T1d;
844 			 Tb = LD(&(ri[WS(is, 28)]), ivs, &(ri[0]));
845 			 Tc = LD(&(ri[WS(is, 12)]), ivs, &(ri[0]));
846 			 Td = VADD(Tb, Tc);
847 			 T1D = VSUB(Tb, Tc);
848 			 T1c = LD(&(ii[WS(is, 28)]), ivs, &(ii[0]));
849 			 T1d = LD(&(ii[WS(is, 12)]), ivs, &(ii[0]));
850 			 T1e = VADD(T1c, T1d);
851 			 T1E = VSUB(T1c, T1d);
852 		    }
853 		    Te = VADD(Ta, Td);
854 		    T1f = VADD(T1b, T1e);
855 		    T50 = VSUB(Td, Ta);
856 		    T4s = VSUB(T1b, T1e);
857 		    {
858 			 V T2U, T2V, T1C, T1F;
859 			 T2U = VSUB(T1D, T1E);
860 			 T2V = VADD(T1B, T1A);
861 			 T2W = VMUL(LDK(KP707106781), VSUB(T2U, T2V));
862 			 T3u = VMUL(LDK(KP707106781), VADD(T2V, T2U));
863 			 T1C = VSUB(T1A, T1B);
864 			 T1F = VADD(T1D, T1E);
865 			 T1G = VMUL(LDK(KP707106781), VSUB(T1C, T1F));
866 			 T3U = VMUL(LDK(KP707106781), VADD(T1C, T1F));
867 		    }
868 	       }
869 	       {
870 		    V Ti, T1L, T1j, T1J, Tl, T1I, T1m, T1M, T1K, T1N;
871 		    {
872 			 V Tg, Th, T1h, T1i;
873 			 Tg = LD(&(ri[WS(is, 2)]), ivs, &(ri[0]));
874 			 Th = LD(&(ri[WS(is, 18)]), ivs, &(ri[0]));
875 			 Ti = VADD(Tg, Th);
876 			 T1L = VSUB(Tg, Th);
877 			 T1h = LD(&(ii[WS(is, 2)]), ivs, &(ii[0]));
878 			 T1i = LD(&(ii[WS(is, 18)]), ivs, &(ii[0]));
879 			 T1j = VADD(T1h, T1i);
880 			 T1J = VSUB(T1h, T1i);
881 		    }
882 		    {
883 			 V Tj, Tk, T1k, T1l;
884 			 Tj = LD(&(ri[WS(is, 10)]), ivs, &(ri[0]));
885 			 Tk = LD(&(ri[WS(is, 26)]), ivs, &(ri[0]));
886 			 Tl = VADD(Tj, Tk);
887 			 T1I = VSUB(Tj, Tk);
888 			 T1k = LD(&(ii[WS(is, 10)]), ivs, &(ii[0]));
889 			 T1l = LD(&(ii[WS(is, 26)]), ivs, &(ii[0]));
890 			 T1m = VADD(T1k, T1l);
891 			 T1M = VSUB(T1k, T1l);
892 		    }
893 		    Tm = VADD(Ti, Tl);
894 		    T1n = VADD(T1j, T1m);
895 		    T1K = VADD(T1I, T1J);
896 		    T1N = VSUB(T1L, T1M);
897 		    T1O = VFNMS(LDK(KP923879532), T1N, VMUL(LDK(KP382683432), T1K));
898 		    T2Z = VFMA(LDK(KP923879532), T1K, VMUL(LDK(KP382683432), T1N));
899 		    {
900 			 V T3w, T3x, T4u, T4v;
901 			 T3w = VSUB(T1J, T1I);
902 			 T3x = VADD(T1L, T1M);
903 			 T3y = VFNMS(LDK(KP382683432), T3x, VMUL(LDK(KP923879532), T3w));
904 			 T3X = VFMA(LDK(KP382683432), T3w, VMUL(LDK(KP923879532), T3x));
905 			 T4u = VSUB(T1j, T1m);
906 			 T4v = VSUB(Ti, Tl);
907 			 T4w = VSUB(T4u, T4v);
908 			 T53 = VADD(T4v, T4u);
909 		    }
910 	       }
911 	       {
912 		    V Tp, T1S, T1q, T1Q, Ts, T1P, T1t, T1T, T1R, T1U;
913 		    {
914 			 V Tn, To, T1o, T1p;
915 			 Tn = LD(&(ri[WS(is, 30)]), ivs, &(ri[0]));
916 			 To = LD(&(ri[WS(is, 14)]), ivs, &(ri[0]));
917 			 Tp = VADD(Tn, To);
918 			 T1S = VSUB(Tn, To);
919 			 T1o = LD(&(ii[WS(is, 30)]), ivs, &(ii[0]));
920 			 T1p = LD(&(ii[WS(is, 14)]), ivs, &(ii[0]));
921 			 T1q = VADD(T1o, T1p);
922 			 T1Q = VSUB(T1o, T1p);
923 		    }
924 		    {
925 			 V Tq, Tr, T1r, T1s;
926 			 Tq = LD(&(ri[WS(is, 6)]), ivs, &(ri[0]));
927 			 Tr = LD(&(ri[WS(is, 22)]), ivs, &(ri[0]));
928 			 Ts = VADD(Tq, Tr);
929 			 T1P = VSUB(Tq, Tr);
930 			 T1r = LD(&(ii[WS(is, 6)]), ivs, &(ii[0]));
931 			 T1s = LD(&(ii[WS(is, 22)]), ivs, &(ii[0]));
932 			 T1t = VADD(T1r, T1s);
933 			 T1T = VSUB(T1r, T1s);
934 		    }
935 		    Tt = VADD(Tp, Ts);
936 		    T1u = VADD(T1q, T1t);
937 		    T1R = VADD(T1P, T1Q);
938 		    T1U = VSUB(T1S, T1T);
939 		    T1V = VFMA(LDK(KP382683432), T1R, VMUL(LDK(KP923879532), T1U));
940 		    T2Y = VFNMS(LDK(KP923879532), T1R, VMUL(LDK(KP382683432), T1U));
941 		    {
942 			 V T3z, T3A, T4x, T4y;
943 			 T3z = VSUB(T1Q, T1P);
944 			 T3A = VADD(T1S, T1T);
945 			 T3B = VFMA(LDK(KP923879532), T3z, VMUL(LDK(KP382683432), T3A));
946 			 T3W = VFNMS(LDK(KP382683432), T3z, VMUL(LDK(KP923879532), T3A));
947 			 T4x = VSUB(Tp, Ts);
948 			 T4y = VSUB(T1q, T1t);
949 			 T4z = VADD(T4x, T4y);
950 			 T52 = VSUB(T4x, T4y);
951 		    }
952 	       }
953 	       {
954 		    V TN, T2p, T2J, T4S, TQ, T2G, T2s, T4T, TU, T2x, T2w, T4O, TX, T2z, T2C;
955 		    V T4P;
956 		    {
957 			 V TL, TM, T2H, T2I;
958 			 TL = LD(&(ri[WS(is, 31)]), ivs, &(ri[WS(is, 1)]));
959 			 TM = LD(&(ri[WS(is, 15)]), ivs, &(ri[WS(is, 1)]));
960 			 TN = VADD(TL, TM);
961 			 T2p = VSUB(TL, TM);
962 			 T2H = LD(&(ii[WS(is, 31)]), ivs, &(ii[WS(is, 1)]));
963 			 T2I = LD(&(ii[WS(is, 15)]), ivs, &(ii[WS(is, 1)]));
964 			 T2J = VSUB(T2H, T2I);
965 			 T4S = VADD(T2H, T2I);
966 		    }
967 		    {
968 			 V TO, TP, T2q, T2r;
969 			 TO = LD(&(ri[WS(is, 7)]), ivs, &(ri[WS(is, 1)]));
970 			 TP = LD(&(ri[WS(is, 23)]), ivs, &(ri[WS(is, 1)]));
971 			 TQ = VADD(TO, TP);
972 			 T2G = VSUB(TO, TP);
973 			 T2q = LD(&(ii[WS(is, 7)]), ivs, &(ii[WS(is, 1)]));
974 			 T2r = LD(&(ii[WS(is, 23)]), ivs, &(ii[WS(is, 1)]));
975 			 T2s = VSUB(T2q, T2r);
976 			 T4T = VADD(T2q, T2r);
977 		    }
978 		    {
979 			 V TS, TT, T2u, T2v;
980 			 TS = LD(&(ri[WS(is, 3)]), ivs, &(ri[WS(is, 1)]));
981 			 TT = LD(&(ri[WS(is, 19)]), ivs, &(ri[WS(is, 1)]));
982 			 TU = VADD(TS, TT);
983 			 T2x = VSUB(TS, TT);
984 			 T2u = LD(&(ii[WS(is, 3)]), ivs, &(ii[WS(is, 1)]));
985 			 T2v = LD(&(ii[WS(is, 19)]), ivs, &(ii[WS(is, 1)]));
986 			 T2w = VSUB(T2u, T2v);
987 			 T4O = VADD(T2u, T2v);
988 		    }
989 		    {
990 			 V TV, TW, T2A, T2B;
991 			 TV = LD(&(ri[WS(is, 27)]), ivs, &(ri[WS(is, 1)]));
992 			 TW = LD(&(ri[WS(is, 11)]), ivs, &(ri[WS(is, 1)]));
993 			 TX = VADD(TV, TW);
994 			 T2z = VSUB(TV, TW);
995 			 T2A = LD(&(ii[WS(is, 27)]), ivs, &(ii[WS(is, 1)]));
996 			 T2B = LD(&(ii[WS(is, 11)]), ivs, &(ii[WS(is, 1)]));
997 			 T2C = VSUB(T2A, T2B);
998 			 T4P = VADD(T2A, T2B);
999 		    }
1000 		    T2t = VSUB(T2p, T2s);
1001 		    T3L = VADD(T2p, T2s);
1002 		    T3O = VSUB(T2J, T2G);
1003 		    T2K = VADD(T2G, T2J);
1004 		    TR = VADD(TN, TQ);
1005 		    TY = VADD(TU, TX);
1006 		    T5F = VSUB(TR, TY);
1007 		    {
1008 			 V T4N, T4Q, T2y, T2D;
1009 			 T5G = VADD(T4S, T4T);
1010 			 T5H = VADD(T4O, T4P);
1011 			 T5I = VSUB(T5G, T5H);
1012 			 T4N = VSUB(TN, TQ);
1013 			 T4Q = VSUB(T4O, T4P);
1014 			 T4R = VSUB(T4N, T4Q);
1015 			 T5j = VADD(T4N, T4Q);
1016 			 T2y = VSUB(T2w, T2x);
1017 			 T2D = VADD(T2z, T2C);
1018 			 T2E = VMUL(LDK(KP707106781), VSUB(T2y, T2D));
1019 			 T3P = VMUL(LDK(KP707106781), VADD(T2y, T2D));
1020 			 {
1021 			      V T4U, T4V, T2L, T2M;
1022 			      T4U = VSUB(T4S, T4T);
1023 			      T4V = VSUB(TX, TU);
1024 			      T4W = VSUB(T4U, T4V);
1025 			      T5k = VADD(T4V, T4U);
1026 			      T2L = VSUB(T2z, T2C);
1027 			      T2M = VADD(T2x, T2w);
1028 			      T2N = VMUL(LDK(KP707106781), VSUB(T2L, T2M));
1029 			      T3M = VMUL(LDK(KP707106781), VADD(T2M, T2L));
1030 			 }
1031 		    }
1032 	       }
1033 	       {
1034 		    V Ty, T2f, T21, T4C, TB, T1Y, T2i, T4D, TF, T28, T2b, T4I, TI, T23, T26;
1035 		    V T4J;
1036 		    {
1037 			 V Tw, Tx, T1Z, T20;
1038 			 Tw = LD(&(ri[WS(is, 1)]), ivs, &(ri[WS(is, 1)]));
1039 			 Tx = LD(&(ri[WS(is, 17)]), ivs, &(ri[WS(is, 1)]));
1040 			 Ty = VADD(Tw, Tx);
1041 			 T2f = VSUB(Tw, Tx);
1042 			 T1Z = LD(&(ii[WS(is, 1)]), ivs, &(ii[WS(is, 1)]));
1043 			 T20 = LD(&(ii[WS(is, 17)]), ivs, &(ii[WS(is, 1)]));
1044 			 T21 = VSUB(T1Z, T20);
1045 			 T4C = VADD(T1Z, T20);
1046 		    }
1047 		    {
1048 			 V Tz, TA, T2g, T2h;
1049 			 Tz = LD(&(ri[WS(is, 9)]), ivs, &(ri[WS(is, 1)]));
1050 			 TA = LD(&(ri[WS(is, 25)]), ivs, &(ri[WS(is, 1)]));
1051 			 TB = VADD(Tz, TA);
1052 			 T1Y = VSUB(Tz, TA);
1053 			 T2g = LD(&(ii[WS(is, 9)]), ivs, &(ii[WS(is, 1)]));
1054 			 T2h = LD(&(ii[WS(is, 25)]), ivs, &(ii[WS(is, 1)]));
1055 			 T2i = VSUB(T2g, T2h);
1056 			 T4D = VADD(T2g, T2h);
1057 		    }
1058 		    {
1059 			 V TD, TE, T29, T2a;
1060 			 TD = LD(&(ri[WS(is, 5)]), ivs, &(ri[WS(is, 1)]));
1061 			 TE = LD(&(ri[WS(is, 21)]), ivs, &(ri[WS(is, 1)]));
1062 			 TF = VADD(TD, TE);
1063 			 T28 = VSUB(TD, TE);
1064 			 T29 = LD(&(ii[WS(is, 5)]), ivs, &(ii[WS(is, 1)]));
1065 			 T2a = LD(&(ii[WS(is, 21)]), ivs, &(ii[WS(is, 1)]));
1066 			 T2b = VSUB(T29, T2a);
1067 			 T4I = VADD(T29, T2a);
1068 		    }
1069 		    {
1070 			 V TG, TH, T24, T25;
1071 			 TG = LD(&(ri[WS(is, 29)]), ivs, &(ri[WS(is, 1)]));
1072 			 TH = LD(&(ri[WS(is, 13)]), ivs, &(ri[WS(is, 1)]));
1073 			 TI = VADD(TG, TH);
1074 			 T23 = VSUB(TG, TH);
1075 			 T24 = LD(&(ii[WS(is, 29)]), ivs, &(ii[WS(is, 1)]));
1076 			 T25 = LD(&(ii[WS(is, 13)]), ivs, &(ii[WS(is, 1)]));
1077 			 T26 = VSUB(T24, T25);
1078 			 T4J = VADD(T24, T25);
1079 		    }
1080 		    T22 = VADD(T1Y, T21);
1081 		    T3E = VADD(T2f, T2i);
1082 		    T3H = VSUB(T21, T1Y);
1083 		    T2j = VSUB(T2f, T2i);
1084 		    TC = VADD(Ty, TB);
1085 		    TJ = VADD(TF, TI);
1086 		    T5A = VSUB(TC, TJ);
1087 		    {
1088 			 V T4E, T4F, T27, T2c;
1089 			 T5B = VADD(T4C, T4D);
1090 			 T5C = VADD(T4I, T4J);
1091 			 T5D = VSUB(T5B, T5C);
1092 			 T4E = VSUB(T4C, T4D);
1093 			 T4F = VSUB(TI, TF);
1094 			 T4G = VSUB(T4E, T4F);
1095 			 T5g = VADD(T4F, T4E);
1096 			 T27 = VSUB(T23, T26);
1097 			 T2c = VADD(T28, T2b);
1098 			 T2d = VMUL(LDK(KP707106781), VSUB(T27, T2c));
1099 			 T3F = VMUL(LDK(KP707106781), VADD(T2c, T27));
1100 			 {
1101 			      V T4H, T4K, T2k, T2l;
1102 			      T4H = VSUB(Ty, TB);
1103 			      T4K = VSUB(T4I, T4J);
1104 			      T4L = VSUB(T4H, T4K);
1105 			      T5h = VADD(T4H, T4K);
1106 			      T2k = VSUB(T2b, T28);
1107 			      T2l = VADD(T23, T26);
1108 			      T2m = VMUL(LDK(KP707106781), VSUB(T2k, T2l));
1109 			      T3I = VMUL(LDK(KP707106781), VADD(T2k, T2l));
1110 			 }
1111 		    }
1112 	       }
1113 	       {
1114 		    V T61, T62, T63, T64, T65, T66, T67, T68, T69, T6a, T6b, T6c, T6d, T6e, T6f;
1115 		    V T6g, T6h, T6i, T6j, T6k, T6l, T6m, T6n, T6o, T6p, T6q, T6r, T6s, T6t, T6u;
1116 		    V T6v, T6w;
1117 		    {
1118 			 V T4B, T57, T5a, T5c, T4Y, T56, T55, T5b;
1119 			 {
1120 			      V T4t, T4A, T58, T59;
1121 			      T4t = VSUB(T4r, T4s);
1122 			      T4A = VMUL(LDK(KP707106781), VSUB(T4w, T4z));
1123 			      T4B = VADD(T4t, T4A);
1124 			      T57 = VSUB(T4t, T4A);
1125 			      T58 = VFNMS(LDK(KP923879532), T4L, VMUL(LDK(KP382683432), T4G));
1126 			      T59 = VFMA(LDK(KP382683432), T4W, VMUL(LDK(KP923879532), T4R));
1127 			      T5a = VSUB(T58, T59);
1128 			      T5c = VADD(T58, T59);
1129 			 }
1130 			 {
1131 			      V T4M, T4X, T51, T54;
1132 			      T4M = VFMA(LDK(KP923879532), T4G, VMUL(LDK(KP382683432), T4L));
1133 			      T4X = VFNMS(LDK(KP923879532), T4W, VMUL(LDK(KP382683432), T4R));
1134 			      T4Y = VADD(T4M, T4X);
1135 			      T56 = VSUB(T4X, T4M);
1136 			      T51 = VSUB(T4Z, T50);
1137 			      T54 = VMUL(LDK(KP707106781), VSUB(T52, T53));
1138 			      T55 = VSUB(T51, T54);
1139 			      T5b = VADD(T51, T54);
1140 			 }
1141 			 T61 = VSUB(T4B, T4Y);
1142 			 STM4(&(ro[22]), T61, ovs, &(ro[0]));
1143 			 T62 = VSUB(T5b, T5c);
1144 			 STM4(&(io[22]), T62, ovs, &(io[0]));
1145 			 T63 = VADD(T4B, T4Y);
1146 			 STM4(&(ro[6]), T63, ovs, &(ro[0]));
1147 			 T64 = VADD(T5b, T5c);
1148 			 STM4(&(io[6]), T64, ovs, &(io[0]));
1149 			 T65 = VSUB(T55, T56);
1150 			 STM4(&(io[30]), T65, ovs, &(io[0]));
1151 			 T66 = VSUB(T57, T5a);
1152 			 STM4(&(ro[30]), T66, ovs, &(ro[0]));
1153 			 T67 = VADD(T55, T56);
1154 			 STM4(&(io[14]), T67, ovs, &(io[0]));
1155 			 T68 = VADD(T57, T5a);
1156 			 STM4(&(ro[14]), T68, ovs, &(ro[0]));
1157 		    }
1158 		    {
1159 			 V T5f, T5r, T5u, T5w, T5m, T5q, T5p, T5v;
1160 			 {
1161 			      V T5d, T5e, T5s, T5t;
1162 			      T5d = VADD(T4r, T4s);
1163 			      T5e = VMUL(LDK(KP707106781), VADD(T53, T52));
1164 			      T5f = VADD(T5d, T5e);
1165 			      T5r = VSUB(T5d, T5e);
1166 			      T5s = VFNMS(LDK(KP382683432), T5h, VMUL(LDK(KP923879532), T5g));
1167 			      T5t = VFMA(LDK(KP923879532), T5k, VMUL(LDK(KP382683432), T5j));
1168 			      T5u = VSUB(T5s, T5t);
1169 			      T5w = VADD(T5s, T5t);
1170 			 }
1171 			 {
1172 			      V T5i, T5l, T5n, T5o;
1173 			      T5i = VFMA(LDK(KP382683432), T5g, VMUL(LDK(KP923879532), T5h));
1174 			      T5l = VFNMS(LDK(KP382683432), T5k, VMUL(LDK(KP923879532), T5j));
1175 			      T5m = VADD(T5i, T5l);
1176 			      T5q = VSUB(T5l, T5i);
1177 			      T5n = VADD(T50, T4Z);
1178 			      T5o = VMUL(LDK(KP707106781), VADD(T4w, T4z));
1179 			      T5p = VSUB(T5n, T5o);
1180 			      T5v = VADD(T5n, T5o);
1181 			 }
1182 			 T69 = VSUB(T5f, T5m);
1183 			 STM4(&(ro[18]), T69, ovs, &(ro[0]));
1184 			 T6a = VSUB(T5v, T5w);
1185 			 STM4(&(io[18]), T6a, ovs, &(io[0]));
1186 			 T6b = VADD(T5f, T5m);
1187 			 STM4(&(ro[2]), T6b, ovs, &(ro[0]));
1188 			 T6c = VADD(T5v, T5w);
1189 			 STM4(&(io[2]), T6c, ovs, &(io[0]));
1190 			 T6d = VSUB(T5p, T5q);
1191 			 STM4(&(io[26]), T6d, ovs, &(io[0]));
1192 			 T6e = VSUB(T5r, T5u);
1193 			 STM4(&(ro[26]), T6e, ovs, &(ro[0]));
1194 			 T6f = VADD(T5p, T5q);
1195 			 STM4(&(io[10]), T6f, ovs, &(io[0]));
1196 			 T6g = VADD(T5r, T5u);
1197 			 STM4(&(ro[10]), T6g, ovs, &(ro[0]));
1198 		    }
1199 		    {
1200 			 V T5z, T5P, T5S, T5U, T5K, T5O, T5N, T5T;
1201 			 {
1202 			      V T5x, T5y, T5Q, T5R;
1203 			      T5x = VSUB(T7, Te);
1204 			      T5y = VSUB(T1n, T1u);
1205 			      T5z = VADD(T5x, T5y);
1206 			      T5P = VSUB(T5x, T5y);
1207 			      T5Q = VSUB(T5D, T5A);
1208 			      T5R = VADD(T5F, T5I);
1209 			      T5S = VMUL(LDK(KP707106781), VSUB(T5Q, T5R));
1210 			      T5U = VMUL(LDK(KP707106781), VADD(T5Q, T5R));
1211 			 }
1212 			 {
1213 			      V T5E, T5J, T5L, T5M;
1214 			      T5E = VADD(T5A, T5D);
1215 			      T5J = VSUB(T5F, T5I);
1216 			      T5K = VMUL(LDK(KP707106781), VADD(T5E, T5J));
1217 			      T5O = VMUL(LDK(KP707106781), VSUB(T5J, T5E));
1218 			      T5L = VSUB(T18, T1f);
1219 			      T5M = VSUB(Tt, Tm);
1220 			      T5N = VSUB(T5L, T5M);
1221 			      T5T = VADD(T5M, T5L);
1222 			 }
1223 			 T6h = VSUB(T5z, T5K);
1224 			 STM4(&(ro[20]), T6h, ovs, &(ro[0]));
1225 			 T6i = VSUB(T5T, T5U);
1226 			 STM4(&(io[20]), T6i, ovs, &(io[0]));
1227 			 T6j = VADD(T5z, T5K);
1228 			 STM4(&(ro[4]), T6j, ovs, &(ro[0]));
1229 			 T6k = VADD(T5T, T5U);
1230 			 STM4(&(io[4]), T6k, ovs, &(io[0]));
1231 			 T6l = VSUB(T5N, T5O);
1232 			 STM4(&(io[28]), T6l, ovs, &(io[0]));
1233 			 T6m = VSUB(T5P, T5S);
1234 			 STM4(&(ro[28]), T6m, ovs, &(ro[0]));
1235 			 T6n = VADD(T5N, T5O);
1236 			 STM4(&(io[12]), T6n, ovs, &(io[0]));
1237 			 T6o = VADD(T5P, T5S);
1238 			 STM4(&(ro[12]), T6o, ovs, &(ro[0]));
1239 		    }
1240 		    {
1241 			 V Tv, T5V, T5Y, T60, T10, T11, T1w, T5Z;
1242 			 {
1243 			      V Tf, Tu, T5W, T5X;
1244 			      Tf = VADD(T7, Te);
1245 			      Tu = VADD(Tm, Tt);
1246 			      Tv = VADD(Tf, Tu);
1247 			      T5V = VSUB(Tf, Tu);
1248 			      T5W = VADD(T5B, T5C);
1249 			      T5X = VADD(T5G, T5H);
1250 			      T5Y = VSUB(T5W, T5X);
1251 			      T60 = VADD(T5W, T5X);
1252 			 }
1253 			 {
1254 			      V TK, TZ, T1g, T1v;
1255 			      TK = VADD(TC, TJ);
1256 			      TZ = VADD(TR, TY);
1257 			      T10 = VADD(TK, TZ);
1258 			      T11 = VSUB(TZ, TK);
1259 			      T1g = VADD(T18, T1f);
1260 			      T1v = VADD(T1n, T1u);
1261 			      T1w = VSUB(T1g, T1v);
1262 			      T5Z = VADD(T1g, T1v);
1263 			 }
1264 			 T6p = VSUB(Tv, T10);
1265 			 STM4(&(ro[16]), T6p, ovs, &(ro[0]));
1266 			 T6q = VSUB(T5Z, T60);
1267 			 STM4(&(io[16]), T6q, ovs, &(io[0]));
1268 			 T6r = VADD(Tv, T10);
1269 			 STM4(&(ro[0]), T6r, ovs, &(ro[0]));
1270 			 T6s = VADD(T5Z, T60);
1271 			 STM4(&(io[0]), T6s, ovs, &(io[0]));
1272 			 T6t = VADD(T11, T1w);
1273 			 STM4(&(io[8]), T6t, ovs, &(io[0]));
1274 			 T6u = VADD(T5V, T5Y);
1275 			 STM4(&(ro[8]), T6u, ovs, &(ro[0]));
1276 			 T6v = VSUB(T1w, T11);
1277 			 STM4(&(io[24]), T6v, ovs, &(io[0]));
1278 			 T6w = VSUB(T5V, T5Y);
1279 			 STM4(&(ro[24]), T6w, ovs, &(ro[0]));
1280 		    }
1281 		    {
1282 			 V T6x, T6y, T6z, T6A, T6B, T6C, T6D, T6E;
1283 			 {
1284 			      V T1X, T33, T31, T37, T2o, T34, T2P, T35;
1285 			      {
1286 				   V T1H, T1W, T2X, T30;
1287 				   T1H = VSUB(T1z, T1G);
1288 				   T1W = VSUB(T1O, T1V);
1289 				   T1X = VADD(T1H, T1W);
1290 				   T33 = VSUB(T1H, T1W);
1291 				   T2X = VSUB(T2T, T2W);
1292 				   T30 = VSUB(T2Y, T2Z);
1293 				   T31 = VSUB(T2X, T30);
1294 				   T37 = VADD(T2X, T30);
1295 			      }
1296 			      {
1297 				   V T2e, T2n, T2F, T2O;
1298 				   T2e = VSUB(T22, T2d);
1299 				   T2n = VSUB(T2j, T2m);
1300 				   T2o = VFMA(LDK(KP980785280), T2e, VMUL(LDK(KP195090322), T2n));
1301 				   T34 = VFNMS(LDK(KP980785280), T2n, VMUL(LDK(KP195090322), T2e));
1302 				   T2F = VSUB(T2t, T2E);
1303 				   T2O = VSUB(T2K, T2N);
1304 				   T2P = VFNMS(LDK(KP980785280), T2O, VMUL(LDK(KP195090322), T2F));
1305 				   T35 = VFMA(LDK(KP195090322), T2O, VMUL(LDK(KP980785280), T2F));
1306 			      }
1307 			      {
1308 				   V T2Q, T38, T32, T36;
1309 				   T2Q = VADD(T2o, T2P);
1310 				   T6x = VSUB(T1X, T2Q);
1311 				   STM4(&(ro[23]), T6x, ovs, &(ro[1]));
1312 				   T6y = VADD(T1X, T2Q);
1313 				   STM4(&(ro[7]), T6y, ovs, &(ro[1]));
1314 				   T38 = VADD(T34, T35);
1315 				   T6z = VSUB(T37, T38);
1316 				   STM4(&(io[23]), T6z, ovs, &(io[1]));
1317 				   T6A = VADD(T37, T38);
1318 				   STM4(&(io[7]), T6A, ovs, &(io[1]));
1319 				   T32 = VSUB(T2P, T2o);
1320 				   T6B = VSUB(T31, T32);
1321 				   STM4(&(io[31]), T6B, ovs, &(io[1]));
1322 				   T6C = VADD(T31, T32);
1323 				   STM4(&(io[15]), T6C, ovs, &(io[1]));
1324 				   T36 = VSUB(T34, T35);
1325 				   T6D = VSUB(T33, T36);
1326 				   STM4(&(ro[31]), T6D, ovs, &(ro[1]));
1327 				   T6E = VADD(T33, T36);
1328 				   STM4(&(ro[15]), T6E, ovs, &(ro[1]));
1329 			      }
1330 			 }
1331 			 {
1332 			      V T3D, T41, T3Z, T45, T3K, T42, T3R, T43;
1333 			      {
1334 				   V T3v, T3C, T3V, T3Y;
1335 				   T3v = VSUB(T3t, T3u);
1336 				   T3C = VSUB(T3y, T3B);
1337 				   T3D = VADD(T3v, T3C);
1338 				   T41 = VSUB(T3v, T3C);
1339 				   T3V = VSUB(T3T, T3U);
1340 				   T3Y = VSUB(T3W, T3X);
1341 				   T3Z = VSUB(T3V, T3Y);
1342 				   T45 = VADD(T3V, T3Y);
1343 			      }
1344 			      {
1345 				   V T3G, T3J, T3N, T3Q;
1346 				   T3G = VSUB(T3E, T3F);
1347 				   T3J = VSUB(T3H, T3I);
1348 				   T3K = VFMA(LDK(KP555570233), T3G, VMUL(LDK(KP831469612), T3J));
1349 				   T42 = VFNMS(LDK(KP831469612), T3G, VMUL(LDK(KP555570233), T3J));
1350 				   T3N = VSUB(T3L, T3M);
1351 				   T3Q = VSUB(T3O, T3P);
1352 				   T3R = VFNMS(LDK(KP831469612), T3Q, VMUL(LDK(KP555570233), T3N));
1353 				   T43 = VFMA(LDK(KP831469612), T3N, VMUL(LDK(KP555570233), T3Q));
1354 			      }
1355 			      {
1356 				   V T3S, T6F, T6G, T46, T6H, T6I;
1357 				   T3S = VADD(T3K, T3R);
1358 				   T6F = VSUB(T3D, T3S);
1359 				   STM4(&(ro[21]), T6F, ovs, &(ro[1]));
1360 				   STN4(&(ro[20]), T6h, T6F, T61, T6x, ovs);
1361 				   T6G = VADD(T3D, T3S);
1362 				   STM4(&(ro[5]), T6G, ovs, &(ro[1]));
1363 				   STN4(&(ro[4]), T6j, T6G, T63, T6y, ovs);
1364 				   T46 = VADD(T42, T43);
1365 				   T6H = VSUB(T45, T46);
1366 				   STM4(&(io[21]), T6H, ovs, &(io[1]));
1367 				   STN4(&(io[20]), T6i, T6H, T62, T6z, ovs);
1368 				   T6I = VADD(T45, T46);
1369 				   STM4(&(io[5]), T6I, ovs, &(io[1]));
1370 				   STN4(&(io[4]), T6k, T6I, T64, T6A, ovs);
1371 			      }
1372 			      {
1373 				   V T40, T6J, T6K, T44, T6L, T6M;
1374 				   T40 = VSUB(T3R, T3K);
1375 				   T6J = VSUB(T3Z, T40);
1376 				   STM4(&(io[29]), T6J, ovs, &(io[1]));
1377 				   STN4(&(io[28]), T6l, T6J, T65, T6B, ovs);
1378 				   T6K = VADD(T3Z, T40);
1379 				   STM4(&(io[13]), T6K, ovs, &(io[1]));
1380 				   STN4(&(io[12]), T6n, T6K, T67, T6C, ovs);
1381 				   T44 = VSUB(T42, T43);
1382 				   T6L = VSUB(T41, T44);
1383 				   STM4(&(ro[29]), T6L, ovs, &(ro[1]));
1384 				   STN4(&(ro[28]), T6m, T6L, T66, T6D, ovs);
1385 				   T6M = VADD(T41, T44);
1386 				   STM4(&(ro[13]), T6M, ovs, &(ro[1]));
1387 				   STN4(&(ro[12]), T6o, T6M, T68, T6E, ovs);
1388 			      }
1389 			 }
1390 		    }
1391 		    {
1392 			 V T6N, T6O, T6P, T6Q, T6R, T6S, T6T, T6U;
1393 			 {
1394 			      V T49, T4l, T4j, T4p, T4c, T4m, T4f, T4n;
1395 			      {
1396 				   V T47, T48, T4h, T4i;
1397 				   T47 = VADD(T3t, T3u);
1398 				   T48 = VADD(T3X, T3W);
1399 				   T49 = VADD(T47, T48);
1400 				   T4l = VSUB(T47, T48);
1401 				   T4h = VADD(T3T, T3U);
1402 				   T4i = VADD(T3y, T3B);
1403 				   T4j = VSUB(T4h, T4i);
1404 				   T4p = VADD(T4h, T4i);
1405 			      }
1406 			      {
1407 				   V T4a, T4b, T4d, T4e;
1408 				   T4a = VADD(T3E, T3F);
1409 				   T4b = VADD(T3H, T3I);
1410 				   T4c = VFMA(LDK(KP980785280), T4a, VMUL(LDK(KP195090322), T4b));
1411 				   T4m = VFNMS(LDK(KP195090322), T4a, VMUL(LDK(KP980785280), T4b));
1412 				   T4d = VADD(T3L, T3M);
1413 				   T4e = VADD(T3O, T3P);
1414 				   T4f = VFNMS(LDK(KP195090322), T4e, VMUL(LDK(KP980785280), T4d));
1415 				   T4n = VFMA(LDK(KP195090322), T4d, VMUL(LDK(KP980785280), T4e));
1416 			      }
1417 			      {
1418 				   V T4g, T4q, T4k, T4o;
1419 				   T4g = VADD(T4c, T4f);
1420 				   T6N = VSUB(T49, T4g);
1421 				   STM4(&(ro[17]), T6N, ovs, &(ro[1]));
1422 				   T6O = VADD(T49, T4g);
1423 				   STM4(&(ro[1]), T6O, ovs, &(ro[1]));
1424 				   T4q = VADD(T4m, T4n);
1425 				   T6P = VSUB(T4p, T4q);
1426 				   STM4(&(io[17]), T6P, ovs, &(io[1]));
1427 				   T6Q = VADD(T4p, T4q);
1428 				   STM4(&(io[1]), T6Q, ovs, &(io[1]));
1429 				   T4k = VSUB(T4f, T4c);
1430 				   T6R = VSUB(T4j, T4k);
1431 				   STM4(&(io[25]), T6R, ovs, &(io[1]));
1432 				   T6S = VADD(T4j, T4k);
1433 				   STM4(&(io[9]), T6S, ovs, &(io[1]));
1434 				   T4o = VSUB(T4m, T4n);
1435 				   T6T = VSUB(T4l, T4o);
1436 				   STM4(&(ro[25]), T6T, ovs, &(ro[1]));
1437 				   T6U = VADD(T4l, T4o);
1438 				   STM4(&(ro[9]), T6U, ovs, &(ro[1]));
1439 			      }
1440 			 }
1441 			 {
1442 			      V T3b, T3n, T3l, T3r, T3e, T3o, T3h, T3p;
1443 			      {
1444 				   V T39, T3a, T3j, T3k;
1445 				   T39 = VADD(T1z, T1G);
1446 				   T3a = VADD(T2Z, T2Y);
1447 				   T3b = VADD(T39, T3a);
1448 				   T3n = VSUB(T39, T3a);
1449 				   T3j = VADD(T2T, T2W);
1450 				   T3k = VADD(T1O, T1V);
1451 				   T3l = VSUB(T3j, T3k);
1452 				   T3r = VADD(T3j, T3k);
1453 			      }
1454 			      {
1455 				   V T3c, T3d, T3f, T3g;
1456 				   T3c = VADD(T22, T2d);
1457 				   T3d = VADD(T2j, T2m);
1458 				   T3e = VFMA(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d));
1459 				   T3o = VFNMS(LDK(KP555570233), T3d, VMUL(LDK(KP831469612), T3c));
1460 				   T3f = VADD(T2t, T2E);
1461 				   T3g = VADD(T2K, T2N);
1462 				   T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f));
1463 				   T3p = VFMA(LDK(KP831469612), T3g, VMUL(LDK(KP555570233), T3f));
1464 			      }
1465 			      {
1466 				   V T3i, T6V, T6W, T3s, T6X, T6Y;
1467 				   T3i = VADD(T3e, T3h);
1468 				   T6V = VSUB(T3b, T3i);
1469 				   STM4(&(ro[19]), T6V, ovs, &(ro[1]));
1470 				   STN4(&(ro[16]), T6p, T6N, T69, T6V, ovs);
1471 				   T6W = VADD(T3b, T3i);
1472 				   STM4(&(ro[3]), T6W, ovs, &(ro[1]));
1473 				   STN4(&(ro[0]), T6r, T6O, T6b, T6W, ovs);
1474 				   T3s = VADD(T3o, T3p);
1475 				   T6X = VSUB(T3r, T3s);
1476 				   STM4(&(io[19]), T6X, ovs, &(io[1]));
1477 				   STN4(&(io[16]), T6q, T6P, T6a, T6X, ovs);
1478 				   T6Y = VADD(T3r, T3s);
1479 				   STM4(&(io[3]), T6Y, ovs, &(io[1]));
1480 				   STN4(&(io[0]), T6s, T6Q, T6c, T6Y, ovs);
1481 			      }
1482 			      {
1483 				   V T3m, T6Z, T70, T3q, T71, T72;
1484 				   T3m = VSUB(T3h, T3e);
1485 				   T6Z = VSUB(T3l, T3m);
1486 				   STM4(&(io[27]), T6Z, ovs, &(io[1]));
1487 				   STN4(&(io[24]), T6v, T6R, T6d, T6Z, ovs);
1488 				   T70 = VADD(T3l, T3m);
1489 				   STM4(&(io[11]), T70, ovs, &(io[1]));
1490 				   STN4(&(io[8]), T6t, T6S, T6f, T70, ovs);
1491 				   T3q = VSUB(T3o, T3p);
1492 				   T71 = VSUB(T3n, T3q);
1493 				   STM4(&(ro[27]), T71, ovs, &(ro[1]));
1494 				   STN4(&(ro[24]), T6w, T6T, T6e, T71, ovs);
1495 				   T72 = VADD(T3n, T3q);
1496 				   STM4(&(ro[11]), T72, ovs, &(ro[1]));
1497 				   STN4(&(ro[8]), T6u, T6U, T6g, T72, ovs);
1498 			      }
1499 			 }
1500 		    }
1501 	       }
1502 	  }
1503      }
1504      VLEAVE();
1505 }
1506 
1507 static const kdft_desc desc = { 32, XSIMD_STRING("n2sv_32"), { 340, 52, 32, 0 }, &GENUS, 0, 1, 0, 0 };
1508 
XSIMD(codelet_n2sv_32)1509 void XSIMD(codelet_n2sv_32) (planner *p) { X(kdft_register) (p, n2sv_32, &desc);
1510 }
1511 
1512 #endif
1513